Llama 3.3 70B Instruct

Checkpoint identity

Official identifier

meta-llama/Llama-3.3-70B-InstructMeta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

Architecture

Dense autoregressive transformer with grouped-query attentionMeta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

Total parameters

70 billionPublished rounded counts; total stored weights determine the preliminary footprint, including all MoE experts.Meta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

Active parameters

Not verified for this recordActive parameters are a compute-related count, not the model's full stored weight count.Meta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

Modalities

text-input, text-outputMeta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

Context

Not verified for this recordNative configuration where directly verified. Extensions, advertised K-labels and configuration differences are preserved in the notes below.Meta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

Release date

2024-12-06Meta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

Download access

Gated: publisher terms / access acceptanceMeta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

License

Llama 3.3 Community LicenseCustom terms include the acceptable-use policy, redistribution notices and Built with Llama attribution, plus naming conditions for distributed models trained using Llama materials or outputs. The release-date 700M monthly-active-user threshold requires a separate Meta license. Review the full terms for a specific use.Meta Llama 3.3 70B Instruct model card and access conditions (external source) · Llama 3.3 Community License (external source)

Theoretical weight-memory footprint

Using the publisher's approximately 70 billion total parameters at 16 bits per parameter gives 140 GB (130.4 GiB) of nominal weights.

This estimate excludes KV cache, activations, quantization metadata, non-weight buffers and serving overhead. It is not a tested GPU fit. MoE routing activates only part of the network per token while the complete expert weights still need storage.

Quantization

Meta documents BF16 and example 8-bit/4-bit loading. Precision reductions require quality and runtime validation; these are not Aguasto benchmark results.

Serving considerations

• 70B is the publisher's approximate total. A separate activated-parameter figure is not stated for this dense model and is left null.
• The card specifies 128k context. This record preserves that published unit without inferring an exact integer from the abbreviated value.
• The selected checkpoint is text only. Hugging Face requires agreeing to the access conditions and sharing contact information. Aguasto has not downloaded, accepted terms, evaluated or deployed it.

Framework documentation

These are third-party documentation starting points. Support depends on the exact checkpoint, runtime version and hardware.